Skip to main content

Adolescent age estimation from magnetic resonance images

Final Report Summary - YOUTH (Adolescent age estimation from magnetic resonance images)

Forensic age estimation of living individuals and human remains has become an important procedure in legal medicine. Example applications include growth predictions for prognostic and therapeutic purposes, diagnosis of endocrinological diseases, victim identification after disasters, assessing asylum seekers entering a country without proper identification documents, or preventing age manipulation in junior-level sports competitions. Being based on bone ossification, the clinically established bone age estimation (BAE) methods employ conventional X-ray examinations of the hand bones to provide means for an objective and reliable age estimation up to 19 years. In cases involving subjects near to the legal majority age, this examination is accompanied by a CT of the clavicle bone, which is one of the last bones to finish ossification, and a panoramic X-ray of the third molar teeth, to enable BAE up to 24 years. A severe drawback of BAE techniques based on standard X-ray or CT imaging is the radiation exposure, which can not be justified for screening healthy children and adolescents, especially in applications of legal or sports medicine. Thus, investigation of BAE methods based on non-invasive magnetic resonance imaging (MRI) has gained in importance, since many countries prohibit scanning involving ionizing radiation without diagnostic reasons. Another benefit of MRI compared to projective 2D X-ray examinations is its volumetric nature, which may provide a foundation for more accurate and reliable BAE. Current methods proposed for BAE in MR images are restricted to best-view cross sections to make use of the same estimation methods that were developed for 2D X-ray images.

The primary aim of the YOUTH project was to perform research on automated age estimation based on the hand bones, clavicle and wisdom teeth in 3D MR images, as well as to design and implement a software tool for computer-assisted age estimation. Thus, the proposed software should increase the accuracy of age estimation compared to state-of-the-art approaches in 2D X-ray and 3D CT images, while removing the harmful exposure to ionizing radiation at the same time. To achieve this goal, the following issues had to be addressed: automatic detection of objects of interest in 3D MR images, automatic segmentation and extraction of features in 3D MR images, and development of a regression algorithm for age estimation.
The Marie Curie Actions fellow Darko Štern has successfully achieved the goal of this project by presenting eight scientific articles and presentations, among which are two publications at the most prestigious conference of the researchers working field, i.e. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2014, and one journal paper in the Annals of Human Biology.
Hence, a fully automated method for localization of the anatomical structures in 3D images was developed and the manuscript “Towards Automatic Bone Age Estimation from MRI: Localization of 3D Anatomical Landmarks” was accepted at the MICCAI 2014 conference for oral presentation, which corresponds to an acceptance rate of below 5%. In order to automatically detect the anatomical structure in MR images, a random regression forest (RRF) approach has been investigated. A two-step multi-scale RRF based approach was developed, that first makes a prediction of the coarse anatomical landmark positions analyzing all shape structures in the image and using feature information from all over the image. This step finds the area, where the landmark locations are expected, and implicitly models the global landmark configuration. Based on these locations, the second step uses more localized information for landmark prediction. Since the proposed algorithm is generic and can be used for localization of arbitrary anatomical structures, it goes beyond the scope of BAE and is of interest for a wide range of medical image analysis tasks, e.g. segmentation, registration, detection and labeling, etc.
A method for determination of legal majority age based on an automated 3D segmentation of the gap between epiphysis and metaphysis of the radius bone was developed. The hypothesis that was tested is that the epiphyseal gap volume of the radius bone, which is segmented using a random classification forest (RCF) from MR hand image training data, predicts the real age using a linear regression model. As a result, the paper “Determination of legal majority age from 3D magnetic resonance images of the radius bone” was presented at the IEEE International Symposium on Biomedical Imaging IEEE-ISBI 2014 and the paper “Legal Majority Age Determination from MR Images of the Radius Bone” at the International Society for Magnetic Resonance in Medicine conference.
As the main contribution, a regression framework was developed, that enables extracting discriminative features from 3D MR images which are relevant for age estimation. The bone age can be estimated from the fusion stages of the epiphyseal gap located between the epiphysis and metaphysis part of the bones. Therefore, a method that extract features that discriminate the fusion stages of the epiphyseal gap based on the known chronological age was developed. Mapping the extracted features to the chronological age can be seen as a regression task, which is modeled using the powerful, yet highly efficient random forest (RF) framework. The decision whether a person is minor or adult is made based on the estimated age. The results were published in “Fully Automatic Bone Age Estimation from Left Hand MR Images” at the MICCAI 2014 conference and in the journal paper for the Annals of Human Biology: “What automated age estimation of hand and wrist MRI data tells us about skeletal maturation in male adolescents”.
The results presented in the six publications are in line with the radiologists performing visual evaluation, i.e. the gold standard in forensic medicine, and with the most prominent automated method trained on a database of a few thousand X-ray images. These results encourage our belief that the project is expected to have a high social and economic impact by providing a novel completely non-invasive and automated method, which could become a new “gold standard” in age estimation. The method may therefore lead to more effective forensic medicine, fairer trials and court decisions. Furthermore, this research could benefit the diagnosis and treatment of metabolic and endocrine disorders that affect growing and development by being able to assist in more accurately assessment of the growth deficits in people with these conditions. During the runtime of this project, collaborations with the Departments of Children Radiology and Orthopaedic Surgery at Medical University of Graz, as well as with the Head of the Biomechanics Lab in the Orthopaedic department of Oslo University Hospital were formed, and a common project proposal for Horizon2020 was written, to more deeply study the applicability of this research for age estimation in orthopaedic surgery.